NVIDIA NCA-GENM study plan

Define a testable hypothesis and one primary metric for a model-change experiment, and write the ship/iterate/abandon decision rule before running anything.

Compute required sample size from a target effect size at alpha 0.05 and power 0.8, then recompute for a halved effect and record how the number changes.

Implement seeded random assignment with fixed train/validation/test splits and prove determinism by regenerating the split twice and diffing the row hashes.

List three leakage risks specific to your setup - target leakage, temporal leakage, duplicate rows across splits - and write an executable check for each.

Build an evaluation harness that reports accuracy, precision, recall, F1, and AUROC on a deliberately imbalanced synthetic dataset.

Run two model variants across five seeds and report mean plus standard deviation for every metric.

Apply a paired bootstrap or paired t-test to the two variants and state explicitly whether the gap is significant at alpha 0.05.

Write a confusion-matrix breakdown identifying which class drives the error and why plain accuracy misleads on this data.

Implement batch gradient descent for logistic regression from scratch in NumPy and confirm it lands within 1% of a scikit-learn baseline on identical data.

Plot learning curves of train versus validation error against training-set size and classify the current fit as underfit, overfit, or well-fit with evidence.

Profile a tabular dataset for missingness, duplicates, outliers, and class imbalance, and record the chosen treatment for each issue.

Compute correlation and mutual information for the top features and flag pairs above 0.8 correlation as redundant.

Build a preprocessing pipeline that pairs images with text, resizes and normalizes images, and tokenizes text into fixed-length tensors.

Assemble a DataLoader with a collate function handling variable-length text and verify tensor shapes by printing three consecutive batches.

Run a modality ablation - image-only, text-only, and fused - and report the metric delta for each configuration.

Document how label alignment and missing-modality samples are handled, covering at least two real broken-sample cases from the data.

Containerize the training entry point with a pinned Dockerfile and reproduce a full run inside the container on a clean machine.

Write unit tests for data transforms and metric functions, reaching at least 80% coverage on those two modules.

Profile one training step with torch.profiler or Nsight Systems and rank the top three time consumers.

Enable automatic mixed precision and measure samples-per-second before and after to quantify the speedup.

Write the same aggregation in SQL and pandas, then reconcile the two outputs row by row and explain any mismatch.

Add data-quality assertions on schema, value ranges, null rates, and duplicate keys so the pipeline fails loudly instead of silently.

Optimize input loading with caching, prefetch workers, and pinned memory, measuring epoch time at 1 versus 4 workers.

Quantize the model to FP16 and INT8 and record the latency, throughput, and accuracy delta for each.

Draft a model card covering intended use, out-of-scope use, training data provenance, evaluation results, and known limitations.

Evaluate the model across at least two subgroups and report the max-minus-min gap on the primary metric.

Write a data-handling note identifying PII fields, retention policy, and license constraints for every dataset used.

Define three drift monitors - input distribution, prediction distribution, and label delay - each with a concrete alert threshold.

Complete two timed 60-question practice sets and score each domain separately against the exam weightings.

Re-derive by hand the sample-size formula, precision/recall/F1, bias-variance decomposition, and AUROC interpretation without notes.

Re-teach your two weakest domains in a written 10-minute explanation each and check them against your week 3 and week 7 deliverables.

Consolidate the capstone repo with a README, exact reproduction commands, and the final results table.